Research & TrendsDeep Dive (Vale)
A gardener in an orange apron lettered Kuaishou prunes a tree of coded paper tags, keeping coin-laden branches while branches with small garment tags fall toward a clothing seller holding a long receipt beside a ledger with a single Revenue column.

Kuaishou Steered Its Ad Generator Toward Revenue and Reported Only Revenue

Kuaishou's GR4AD-X paper found that its ad generator's likeliest ads and its highest-eCPM ads barely overlapped, steered generation toward eCPM, and credits the full system with a 3.07% revenue gain. Its tables report revenue and serving speed but no click, conversion or advertiser-return figure, although most of the same researchers published conversion and small-advertiser figures for GR4AD in February.

Neritus Vale

Kuaishou’s ad engineers compared the ads their generator rated likeliest with those their ranking stage expected to earn most, and found the lists barely touched. In GR4AD-X, submitted to arXiv on 28 October by a Kuaishou Technology team, the overlap is “only approximately 0.02”. The fix steers generation toward the ads that pay; its online evidence then shows what Kuaishou earned and not what viewers or advertisers got.

A generative recommender has to spell an ad before it can show it, as we reported in October. Each of the paper’s 7.556 million ads carries a short code called a semantic ID. The model writes it token by token, and a beam search drops all but the likeliest partial codes. The paper warns that an underestimated prefix “may cause a high-eCPM advertisement to be pruned before its complete SID is generated.”

The paper’s diagnosis takes one sentence: “Next-token prediction fits logged interactions, whereas advertising systems optimize commercial value such as eCPM.” In ad-tech usage, eCPM is the revenue an ad is expected to earn per thousand impressions. The generator learned what viewers did; Kuaishou is paid on something else.

GR4AD-X feeds value in before, during and after generation. A training step, PV-RSPO, passes eCPM rewards up the tree of partial codes, so each prefix inherits the average reward of the sampled ads beneath it. PV-BeamSearch then ranks a widened pool by probability and by predicted value and fuses the two rankings. A final reranking step re-sorts the finished candidates using “eCPM-related signals”.

The generator now ranks far more like the revenue ranker. On logged requests, top-10 HitRate against eCPM rises from 0.021 to 0.298 before reranking, compared with the same model without value alignment. The beam has a set width, so every prefix that value lifts in pushes out one that probability alone would have kept. The paper does not say which ads were pushed out, or whether viewers missed them.

Revenue is where the paper keeps score. Its A/B table adds components in sequence to the deployed GR4AD system; the full stack ends 3.07% ahead. Most of that lift arrives before the new value steps. Multi-path IDs, which also help ads survive pruning, and model scaling account for 2.00 points.

Value alignment supplies the other 1.07 points. Value-guided beam search, the step aimed at the diagnosed pruning failure, adds 0.21, the least of any step. It was added after PV-RSPO, so its share depends on run order.

The largest value step comes after the beam has closed. Reranking the finished list adds 0.58 points. It rescues no pruned ad. PV-RSPO, which teaches the generator to value prefixes during training, adds 0.28.

The test protocol covers one metric. Each variant ran on a randomly assigned 5% of traffic for seven days. The significance claim covers “all reported revenue improvements”, which is every improvement the online tests report.

The paper reports no click-through rate, conversion rate, advertiser return or user-experience guardrail. Yet it ties the order and value of candidates to “platform revenue and the fulfillment of advertiser objectives”. Only the first is measured. The only cost it reports is Kuaishou’s own: value alignment keeps 75% of the base model’s serving throughput.

Most of the same researchers published user- and advertiser-side figures in February. GR4AD, the paper behind the baseline, shares eight of GR4AD-X’s authors and carries a section headed “Business Indicators”. It credited “more precise interest modeling” with a 10.17% improvement in ad conversion rates.

GR4AD also reported 17.5% more ad delivery for small and medium-sized advertisers, and claimed “a win-win-win outcome for the platform, advertisers, and users”. GR4AD-X reports the first win only.

Alibaba’s engineers printed the other two lines for a related fix. Their BAR paper targets retrieval that cannot see the live bids its rankers sort on. In the first version, the online test reported click-through rate “to evaluate user experience” and ROI, advertisers’ GMV over platform revenue, “to measure the advertiser’s experience”. Platform revenue rose 4.32% while ROI held flat, so advertisers’ sales kept pace with their bills.

Kuaishou’s investor language is broader than the paper’s evidence. Its second-quarter release quoted chief executive Cheng Yixiao saying AI was “enhancing both user experience and commercial efficiency”, a sentence that spans content, online marketing, e-commerce and internal operations. The same release said the company would extend its generative recommendation and intelligent bidding models to live streaming, search and pan-shelf e-commerce, “unlocking marketing budgets from our clients”. On an ad platform, revenue is what advertisers are billed, so a revenue table also records what they paid.

For clothing sellers, the missing lines are the ones they buy on. At a September 2025 salon, Kuaishou’s marketing arm pitched its paid-traffic tools to apparel merchants for the September-to-February peak, according to a write-up on Digitaling originally credited to 财经社. Kuaishou’s Net Transaction ROI bidding product, which the company says helps merchants reduce return rates, reached 55% client penetration in the second quarter. A seller bidding on that product is watching advertiser-side outcomes that GR4AD-X never prints.

The strongest defence is that eCPM already contains the viewer. Alibaba’s BAR abstract defines eCPM as predicted click-through rate times the bid, so on Alibaba’s display ads an ad nobody clicks scores low at any price. GR4AD-X never defines Kuaishou’s eCPM, so the defence assumes it works the same way. The reranker also has a behavior head that models “the progressively deeper conversion chain”. For the thesis to fail, the extra revenue would have to come from viewers responding more rather than from higher bidders displacing lower ones.

The paper gives no way to check that defence. A predicted click is Kuaishou’s estimate, folded into Kuaishou’s score; a measured click is evidence, and GR4AD-X prints none. The behavior head shares the final vote with the eCPM head, and its effect on outcomes goes unreported. The figures that would settle it sit in the A/B arms that produced the 3.07%.

GR4AD-X aligned the generator with value and then measured only Kuaishou’s.

If Kuaishou keeps publishing revenue alone, merchants should treat each value step as a change in what they pay until the platform shows click-through, conversion and ROI from the same arms. Which ledger to publish is Kuaishou’s choice. Whether to buy peak-season traffic without seeing the advertiser’s ledger is the merchant’s.